ctt artifacts 2026-07-02 workspace/docs/cil_format.md

workspace/docs/cil_format.md

@@ -43,10 +43,15 @@ ranking, CAR, and oracle computations.
 
 Canonical metric API:
 
-- `cil/metrics.py` defines Branch CAR, PTR@K, SelectorRegret@K, SupportGap,
-  support/selector CAR decomposition, negative-near rate,
+- `cil/metrics.py` defines BranchCAR, measured OutcomePTR@K,
+  SelectorRegret@K, measured SupportGap, support/selector CAR decomposition,
+  ProxyPositiveTangentCoverage@K (PPTC@K), negative-near rate,
   positives-closer-than-negatives, pairwise causal dominance ECE, and
   micro/task-macro/seed-macro bootstrap summaries.
+- Distance-only support diagnostics must be reported as PPTC, never as PTR or
+  OutcomePTR. `scripts/eval_metrics.py --mode measured` refuses rows without
+  `candidates_evaluated=true`; `--mode proxy` exports PPTC/NegativeNear/proxy
+  support distance.
 - `runs/reproduce_v0/metrics.json` is the current Phase-0 provenance bundle:
   git hash, data hash, split hash, input file hashes, gate status, support
   proxy rows, and leakage-audit summary.
@@ -68,15 +73,18 @@ Canonical branch families:
 - `negative_antigoal`
 - `learned_generator`
 
-Generator baselines:
+Generator baselines and CTT:
 
 - V0 transported residual retrieval copies train-state tangents into the current
   state and lets the learned utility field select among them.
 - V1 utility-weighted residual retrieval changes proposal support before
   selection by weighting train tangents with retrieval affinity and measured
   source advantage, `exp(-distance / tau + rho * delta_utility)`.
-- The full CIL-Atlas generator should learn positive causal tangent support
-  directly, with V0/V1 kept as diagnostic baselines rather than the main novelty.
+- Causal Tangent Transport (CTT) is the current method spine:
+  `T_phi(z_source, z_target, xi_source_positive) -> xi_target_positive`.
+  CTT starts from measured train positive tangents and transports them into the
+  target chart. V0/V1/CVAE/flow rows remain diagnostic baselines rather than the
+  main novelty.
 
 Learned generator targets:
 
@@ -86,11 +94,12 @@ Learned generator targets:
   `data/cil_charts/{split}/charts_*.npz` plus `index.json`. The train export
   is the only retrieval/generator-training index; non-train exports must be
   evaluator-only.
-- `scripts/audit_leakage.py` checks chart indexes for split misuse,
+- `scripts/audit_cil_charts.py` checks chart indexes for split misuse,
   non-train outcome exposure, missing shards, and same-state group/state-hash
-  overlap between train and eval splits. The current train chart DB contains
-  2,873 charts and 45,968 rows including base branches, and its audit passes
-  with zero violations.
+  overlap between train and eval splits. The current split export contains
+  2,873 charts and 45,968 rows including base branches. The train retrieval DB
+  exposes 2,044 charts and 32,704 rows; validation/test are evaluator-only.
+  The current audit passes with zero violations.
 - The base priority is deployment-clean (`policy`, `policy_residual`, `anchor`,
   `base`) with `expert` as a final fallback for current train-only CIL shards.
 - Labels are measured from same-state utility contrast:
@@ -102,9 +111,9 @@ Learned generator targets:
   object-centric spline tangent space.
 - `scripts/eval_positive_tangent_memory.py` is a leakage-checked diagnostic
   baseline: train-only positive tangents become diverse task-level prototypes,
-  and heldout groups report PTR proxy, negative-near rate, and whether the
-  nearest proposal is closer to a hidden positive than to hidden negatives. It
-  should be reported as support evidence, not as the final learned generator.
+  and heldout groups report PPTC, negative-near rate, and whether the nearest
+  proposal is closer to a hidden positive than to hidden negatives. It should be
+  reported as support evidence, not as the final learned generator.
 - `scripts/eval_positive_tangent_local_atlas.py` evaluates local chart-neighborhood
   reuse: for each heldout chart it retrieves train-only positive tangents from
   nearby observation-language-task charts. This is not the final generator, but
@@ -112,22 +121,22 @@ Learned generator targets:
   also score candidates by distance from local train negatives, which is a
   diagnostic for whether negative boundaries alone can replace local positive
   support. Current K16 support-proxy results say no: pool16 local positives
-  reach 23.66% PTR at RMS<=0.20 and 52.69% at RMS<=0.40, while pool64
-  negative-margin reranking keeps or lowers strict PTR and does not reduce the
+  reach 23.66% PPTC at RMS<=0.20 and 52.69% at RMS<=0.40, while pool64
+  negative-margin reranking keeps or lowers strict PPTC and does not reduce the
   5.33% strict negative-near rate.
 - `scripts/eval_positive_tangent_chart_synthesis.py` is the next local-chart
   diagnostic. It keeps train-only positive chart atoms and adds barycentric
   means over nearby chart neighborhoods, testing whether positive support is
   better expressed as local chart coordinates than as raw prototype replay.
   Current best keeps 15 direct local atoms plus one 16-neighbor chart mean,
-  preserving K16 PTR at 23.66% / 52.69% and strict negative-near at 5.33% while
+  preserving K16 PPTC at 23.66% / 52.69% and strict negative-near at 5.33% while
   improving positives-closer-than-negatives to 65.33%. This remains an offline
   support proxy; successful settings should motivate the learned object-centric
   atlas generator rather than become the final deployment method by themselves.
 - `scripts/train_positive_tangent_cvae.py` trains a first raw-action CVAE over
   train-only positive tangents. The companion
   `scripts/summarize_positive_tangent_cvae_sweep.py` ranks temperature/beta
-  sweeps by heldout PTR and negative-near rates. This baseline is useful for
+  sweeps by heldout PPTC and negative-near rates. This baseline is useful for
   falsifying raw action-chunk likelihood as the final geometry.
 - `scripts/train_positive_tangent_spline_cvae.py` trains a keyframe-spline CVAE
   over start/midpoint/endpoint residual codes and decodes samples back into